Medium- and long-run performance comparison of LLM-based and Q-learning pricing agents

Compare the relative performance in the medium and long run of GPT-4-based pricing agents using prompt prefixes P1 and P2 and Q-learning pricing agents in the repeated Bertrand duopoly environment with logit demand, including outcomes such as prices and profits.

Background

In experiments where an LLM-based agent competes against a Q-learning agent, the Q-learning agent is in its exploration phase over the 300-period horizon, leading to short-run profit differences favoring the LLM-based agents.

The authors explicitly defer the systematic comparison of relative performance across these agent types to future research, specifically over medium- and long-run horizons, to understand how learning dynamics and eventual convergence affect profits and pricing behavior.